MIKA: Source reconstructibility for robust document processing with LLMs: Verification and domain partitioning in enterprise settings

Large language models (LLMs) in enterprise document processing fail in two structurally distinct ways: value hallucination on precision entities (a single wrong digit in a dosage or tax code) and cross-domain contamination in retrieval-augmented generation (RAG). We present MIKA , a distributed document-processing system with two contributions. VERA , a multi-layer post-generation verifier, anchors each extracted claim to the specific source region it came from, raising no false positives over ≈ 140,000 LLM-extracted precision fields (real extractions over synthetic clean documents; 0 / ≈ 140,000 , Wilson 95% upper bound ≈ 2.7 × 1 0 − 5 ) and detecting all 2800 injected precision errors under distinct field labels on a reproducible synthetic benchmark (Wilson 95% lower bound ≥ 0.9986 ). VERA verifies fidelity to the deterministic OCR/parser source , not real-world correctness: an OCR error the LLM reproduces faithfully passes as valid. Detected numeric errors are persisted as queryable anti-patterns in a Numeric Determinism Layer (NDL). PC-RAG (Partitioned-Context RAG) enforces an index-level partitioning invariant, K i ∩ K j = 0̸ . In a two-arm pilot on a real balanced multi-domain corpus (production OCR and embeddings), partitioning reduces measured cross-domain retrieval contamination from 3.3% under a unified index to 0%; the 0% is a structural consequence of the invariant under correct routing. A centroid router attains 96.25% accuracy (in-sample pilot estimate), bounding end-to-end retrieval contamination at ≈ 3.75 % for this pilot. Both contributions realize one principle, preservation of source reconstructibility , supporting verification and audit workflows. Throughput (2644 pages/min, 69.9 × speedup), established in separate work, is quoted as a substrate property. Validation is on Spanish-language extractive tasks; generalization to other languages and semi-structured sources is argued structurally, not demonstrated. We report MIKA as a scoped proof-of-concept : the evidence base is a reproducible synthetic benchmark, a 104-document real-OCR triage run, and a 124-document retrieval pilot, not an enterprise-scale validation.